View PDF

Download full issue

Cell
Volume 186, Issue 1, 5 January 2023, Pages 178-193.e15

Theory

An approximate line attractor in

the hypothalamus encodes an
aggressive state
Aditya Nair 1, 2, 3… Ann Kennedy 1, 2, 3, 8

Show more

Outline Share Cite

https://doi.org/10.1016/j.cell.2022.11.027
Get rights and content
Under a Creative Commons license Open access

Highlights
• Dynamical system analysis reveals a
line attractor in VMHvlEsr1 neuronal
activity

• The line attractor represents a

neural correlate of aggressive
escalation

• Individual differences in aggression

are correlated with line attractor
properties

• MPOAEsr1 utilizes rotational not line

attractor dynamics to encode mating
behavior

Summary
The hypothalamus regulates innate social
behaviors, including mating and aggression.
These behaviors can be evoked by optogenetic
stimulation of specific neuronal subpopulations
within MPOA and VMHvl, respectively. Here, we
perform dynamical systems modeling of
population neuronal activity in these nuclei
during social behaviors. In VMHvl, unsupervised
analysis identified a dominant dimension of
neural activity with a large time constant (>50 s),
generating an approximate line attractor in
neural state space. Progression of the neural
trajectory along this attractor was correlated with
an escalation of agonistic behavior, suggesting
that it may encode a scalable state of
aggressiveness. Consistent with this, individual
differences in the magnitude of the integration
dimension time constant were strongly
correlated with differences in aggressiveness. In
contrast, approximate line attractors were not
observed in MPOA during mating; instead,
neurons with fast dynamics were tuned to
specific actions. Thus, different hypothalamic
nuclei employ distinct neural population codes
to represent similar social behaviors.

Graphical abstract

Download : Download high-res image (203KB)

Download : Download full-size image

Previous Next

Keywords
aggression; courtship; hypothalamus; VMH;
MPOA; line attractor; dynamical systems;
innate behavior; calcium imaging; rSLDS

Introduction
A fundamental problem in neuroscience is to
understand how the brain controls innate
behaviors. Many such behaviors are governed by
the hypothalamus, a deep subcortical brain
region present in all vertebrates.1,2 Classical
brain stimulation and lesion experiments have
implicated different hypothalamic regions
(“nuclei”) in diverse innate behaviors (reviewed in
Paredes and Baum,3 Siegel et al.,4 Canteras,5
King,6 Kruk,7 Swanson,8 and Simerly9). More
recently, optogenetic stimulation has identified
genetically marked neuronal subpopulations that
can evoke such behaviors10,11,12,13 (reviewed in
Yamaguchi,14 Zha and Xu,15 Augustine et al.,16
and Sternson17). Genetic ablation or reversible
silencing has demonstrated that these
subpopulations are essential for natural
occurrences of these behaviors.10,11,12,18

An important open question is how the activity

of these neural subpopulations during naturally
occurring behavior reflects their “causative”
function. Relatively few single-unit recordings
have been performed in hypothalamic nuclei
because of their inaccessibility.13,19,20,21
Recordings of bulk calcium signals22 have
confirmed that these neuronal subpopulations
are active during the natural behaviors they can
artificially evoke.23,24,25 However, this averaging
method obscures individual cell activity patterns.

Miniature head-mounted microscopes allow

calcium imaging with single-cell resolution in
freely moving animals.26,27 Application of this
approach to the hypothalamus has identified
cells exhibiting stimulus-locked activity during
natural behavior.28,29,30 For example, imaging of
estrogen receptor type 1 (Esr1)-expressing
neurons in the medial preoptic area (MPOA),
whose optogenetic activation can elicit mounting
behavior in male mice,31,32 has revealed cells that
respond specifically during spontaneous
mounting of females (see also Figure 1E). Such
results, together with single-cell transcriptomic
analysis, have reinforced the prevailing view that
the hypothalamus controls different survival
behaviors via genetically determined,
functionally specific neuronal
subpopulations.33,34

Download : Download high-res image (2MB)

Download : Download full-size image

Figure 1. Cytoarchitectures and cellular

representations in a neural system regulating
social behavior

(A and B) Cytoarchitecture and input-output

maps of MPOA33,35 (A) and VMHvl31,36,37,38 (B).

(C and D) Example traces from Esr1+ neurons in

MPOA (C) and VMHvl (D).

(E and F) Clustering of recorded Esr1+ neurons in

MPOA (E, n = 306 neurons from 3 mice) and
VMHvl (F, n = 391 neurons from 4 mice) using a
regression model. Rows, hand-annotated
behaviors; columns, individual neurons.

The case of aggression, however, presents a

paradox seemingly at odds with this view. On one
hand, optogenetic stimulation of Esr1+ neurons
in the ventrolateral subdivision of the
ventromedial hypothalamus (VMHvl) neurons
triggers attack behavior,12,39,40,41 identifying these
neurons as the likely cellular substrate of
electrical brain-stimulated aggression.4,7,42
Conversely, genetic ablation of VMHvl neurons
expressing the progesterone receptor (PR; co-
expressed with Esr1) or optogenetic silencing of
VMHvlEsr1 neurons blocks natural aggression12,18

On the other hand, miniscope imaging of

VMHvlEsr1 neurons during natural fighting
revealed surprisingly few cells that exhibited
time-locked, attack-specific activity.29 Instead,
most such neurons exhibited “mixed selectivity,”
responding during different phases of an
aggressive interaction. Different subsets of Esr1+
neurons responded to male versus female
conspecifics, suggesting an encoding of
conspecific sex.29,31,43 Nevertheless, decoders
trained on VMHvlEsr1 neural imaging data could
accurately distinguish episodes of attack from
sniffing.29

Thus, observational versus perturbational studies

of VMHvlEsr1 neurons yield seemingly
inconsistent views: these neurons causally
control aggressive behavior; however, very few of
them are specifically “tuned” to attack. There are
two possible explanations for this paradox. First,
the small fraction of VMHvlEsr1 neurons that are
more active during attack may be the ones
responsible for the specific causative influence of
this population. Alternatively, the majority of
VMHvlEsr1 neurons, despite their mixed
behavioral selectivity, may control attack through
some type of population code.

In other systems where there is no clear

correlation between single-unit spiking patterns
and behavior, modeling neural populations as a
dynamical system44,45,46 (reviewed in Vyas et al.47)
has revealed signals in the dynamics of
population activity that can robustly predict
motor actions.48,49 We have therefore carried out
similar modeling of VMHvlEsr1 neural activity
dynamics during naturalistic social behaviors,
using legacy data from previous studies.29,31,43
Our results reveal line attractor dynamics
in VMHvl that correlate with escalating levels of
aggressive behavior, suggesting that they may
represent or encode an aggressive internal state.
Strikingly, line attractor dynamics are absent in
MPOA activity during both mating and
aggression. This analysis therefore reveals
fundamental differences in the neural coding of
social behaviors by different hypothalamic
nuclei.

Results

Cellular tuning analysis confirms

behaviorally selective neural populations in
MPOA but not in VMHvl
Calcium imaging of MPOAEsr1 or VMHvlEsr1
neurons revealed distinct patterns of neuronal
activation during social interactions29,31
(Figures 1A and 1B). To quantify these
differences, we re-analyzed calcium imaging
data31 from sexually experienced male
C57Bl/6NEsr1-2A-Cre/+ mice during standard
resident-intruder assays, using male or female
BalbC intruders (Figures 1C and 1D). We then
computed the mean activity of each neuron
during each of 14 different hand-annotated
actions and clustered them using a regression
model (VMHvl: n = 306 neurons from 3 mice;
MPOA: n = 391 neurons from 4 mice, see STAR
Methods).

Confirming previous observations,29,31 many

MPOA clusters contained neurons only active
during specific behavioral actions, such as
intromission or mounting toward females
(Figure 1E). In contrast, most VMHvlEsr1 neurons
were activated in response to either males or
females, with very few neurons showing
behavior-specific activation (Figure 1F).

Unsupervised dynamical systems analysis

of neural activity during social behavior
In other systems, population analysis via fit
dynamical systems has revealed a neural
encoding of behavioral actions that were not
apparent in neuron-by-neuron analysis.47,48,50,51
We therefore investigated whether behavioral
representations among VMHvlEsr1 neurons
might be encoded at a population level, using an
unsupervised dynamical systems approach.

To do so, we fit a dynamical model to the

population activity of VMHvlEsr1 cells from each
of multiple mice (n = 6), from two different
studies31,43 in which recordings were made
throughout male-male or male-female
encounters (average duration 5.1 ± 0.68 min and
11.4 ± 0.68 min, respectively; mean ± SEM).
Specifically, we fit a recurrent switching linear
dynamical system (rSLDS) model,52 which
approximates a complex non-linear dynamical
system as a composite of more easily
interpretable linear dynamical systems, or
“states” (Figure S1A).

Download : Download high-res image (2MB)

Download : Download full-size image

Figure S1. Unsupervised discovery of aggression-

enriched states in VMHvl, related to Figure 2

(A) Types of neural states identified by rSLDS. B1,

B2: behaviors; Q0, Q1: periods of quiescence
between behavior bouts; S0, S1, S2: rSLDS states.
Case 1: rSLDS states cannot distinguish behavior
versus internal states. Case 2: rSLDS reflects
internal state-encoding due to persistence during
behavioral quiescence.

(B) Optimization of number of rSLDS states in

example VMHvl mouse 1. Model performance is
measured as ELBO (see STAR Methods).

(C) Same as (B), but for dimensionality.

(D) Variance explained by dimension chosen in

(C).

(E) Convergence of model performance.

(F) Creation of a bounded model performance

metric (forward simulation error [FSE]; see STAR
Methods).

(G) FSE for VMHvl mouse 1 & 2.

(H) Average model performance (FSE) before and

after training (n = 6 mice,∗∗∗p < 0.001)

(I) (I1) rSLDS states in VMHvl mouse 1. (I2)

Comparison of rSLDS states with behaviors. (I3)
Behavioral composition of rSLDS states. State 3
possesses the highest amount of attack behavior
across mice (see J and K). (I4) Probability of attack
aligned to the onset of state 3 (n = 6 mice). (I5)
Timescale of behavior bouts and discovered
states epochs. (I6) State transition diagram from
empirical transition probabilities.

(J) Same as F2, F3, F5 but for VMHvl mouse 2.

(K) Same as F2, F3, F5 but for VMHvl mouse 3.

rSLDS first reduces neural activity to a set of

latent variables (also called “dimensions” or
“factors”), defining a low-dimensional “state
space” in which the time-evolving population
neural activity vector can be analyzed
(Figure 2A➀). Population activity in this low-
dimensional space is then segmented into a set
of discrete states (Figure 2A②) while fitting a
linear dynamical system model (Figure 2A③) to
neural activity within each state. Each state has a
different dynamics matrix, which dictates how
neural activity evolves over time from any given
point within that state space. Quantitative
examination of parameters from this matrix after
model fitting can unveil dynamical properties of
the neural circuit, such as the time constant of
each dimension.46 Finally, to visualize more
easily the dynamical properties of each state, we
plotted its “flow field” in 2D using principal
component analysis (PCA) (Figure 2A④ right; see
STAR Methods).

Download : Download high-res image (1MB)

Download : Download full-size image

Figure 2. Dynamical analysis of VMHvl neural

activity reveals an integrator dimension that
correlates with aggressive escalation

(A) Schematic illustrating rSLDS52 analysis. Steps

➀–④ are shown sequentially for illustrative
purposes only.

(B) Time constants of rSLDS dimensions (see A➀)

in attack enriched state from VMHvl mouse 1.
Dimensions with longest (red dot) and shortest
(yellow dot) time constants are indicated.

(C) Projection onto time axis of integration

dimension with overlayed behavior annotations.

(D) Average time constant of all dimensions,

arranged in decreasing order. (∗∗∗p < 0.001, n = 6
mice.)

(E) Average F1 score of binary decoder of

behavior pairs trained on integration dimension
activity (∗∗p < 0.005, ∗p < 0.01, n = 6 mice).

(F) Cumulative distribution of integration

dimension value (normalized) for different
behaviors.

(G) Projection of fastest dimension in example

VMHvl mouse 1.

(H) Performance of binary decoder of behavior

pairs trained on fastest dimension activity (n = 6
mice). For additional data, see Figures S1, S2, and
S3.

In fitting the rSLDS model, we chose the

minimum number of states and dimensions that
could capture 90% of observed variance in neural
activity, determined using cross validation in
each mouse separately (Figures S1B–S1E; 7–8
dimensions [7.2 ± 0.1, N = 6 mice] and 3–4 states).
We evaluated the “goodness of fit” of each model
iteration using both the log likelihood of the
data52 and an additional metric that we call the
“forward simulation error” (FSE; Figure S1F; see
STAR Methods). Plotting the FSE over time
allows visualization of periods wherein model
performance drops (Figure S1G). By this metric,
our best-fit models captured most of the variance
in neural data (model performance [1-FSE] =
0.72 ± 0.02, N = 6 mice; Figure S1H).

The rSLDS framework allows the fit dynamical

system models to be either autonomous or to
receive external input. Since VMHvl neuron
firing rates correlate with the distance to another
male or to male mouse urine,53 likely reflecting
the concentration of chemosensory cues,54 we
used the distance between animals and their
facing angle as a proxy for external sensory input
strength53,55 (see STAR Methods).

rSLDS analysis of VMHvl neural activity

discovers an integration dimension that
correlates with aggressive escalation
Next, we performed retrospective alignment of
the unsupervised neural data model with
behavioral annotations over time. This
comparison revealed that the probability of
attack was elevated during a single rSLDS state
(state 3, Figures S1I–S1K). Importantly, attacks
were not time-locked to the onset/offset of this
state; rather, epochs of this state outlasted
individual attack bouts (state 3 epoch duration:
79.5 ± 5.5 s, attack bout duration: 4.86 ± 0.44 s,
N = 6 mice, Figures S1I5, S1J3, and S1K3). This
suggests that the state did not simply represent
motor activity (Figure S1A, cf. case 2 versus 1).

To better understand the neural population

dynamics related to attack behavior, we examined
the dynamics matrix for this state, which
describes how dimensionally reduced neural
activity in that state changes over time. The
eigenvalues of this matrix reflect the rate at
which activity along each of these dimensions
decays to zero following external input and can
be converted to a time constant for each
dimension.56,57 Input to dimensions with short
time constants will quickly decay to zero, whereas
input to dimensions with long (large) time
constants persists and decays slowly. Strikingly,
one of the rSLDS dimensions had an estimated
time constant of over 100 s that was significantly
higher than that of all other dimensions
(Figures 2B red dot, 2C, and 2D, N = 6 mice).
Because systems with long time constants
approximately integrate their input over time, we
refer to the longest time constant dimension as
the “integration” dimension.58,59

The integration dimension accounted for

19.5% ± 1.9% of the overall variance in neural
activity (N = 6 mice). In contrast, a support vector
machine (SVM) decoder trained on all neural
data to distinguish attack from sniffing periods
explained much less variance (0.3% ± 0.1%, N = 6
mice, p < 0.001, Figure S2B). Examining the
activity of individual neurons that were weighted
strongly in the integration dimension
(Figure S2D) revealed that around 20% of
neurons per animal contributed to this
dimension, with some showing ramping and
persistent activity (Figures S2I, S2J, S2L, and
S2N). Moreover, most of these neurons were
tuned to male intruders (Figures S3A and S3B).
Thus, the integration dimension encapsulates a
signal that is present at the level of at least some
individual neurons but is also an emergent
property of the population.51

Download : Download high-res image (2MB)

Download : Download full-size image

Figure S2. Characterization of aggression-

integration dimension, related to Figure 2

(A) Variance explained by a generalized linear

model trained to predict integration dimension
from pose features including distance between
mice, facing angle, speed, acceleration, and
velocity of resident mouse (mean: 0.28 ± 0.04 R2,
n = 6 mice).

(B) Fraction of overall variance explained by

integration dimension (purple) compared to
variance explained by decoder dimension trained
to distinguish attack from sniff bouts
(integration dimension mean: 19.5% ± 1.9%,
attack decoder mean: 0.3% ± 0.1%, n = 6 mice,
∗∗∗p < 0.001).

(C) Decoding behaviors from non-integration

dimensions (average across dimensions, n = 6
mice).

(D) Absolute rSLDS weight on integration

dimension of VMHvl mouse 1 (cell number on x
axis), sorted by choice probability values for male
versus female intruder encounter.

(E–G) Paradigm to account for spurious

correlations: decoding threshold obtained using
integration dimension of mouse 1 (E, purple line)
is used on integration dimension from mouse 2
(F). Spurious correlations lead to low F1 scores
(F) while true correlations retain high F1 scores
(G).

(H) Decoding behaviors using paradigm

described above (∗∗p < 0.005, n = 6 mice).

(I) Normalized activity of neurons times rSLDS

weight for cells with significant weights for
integration dimension of VMHvl mouse 1.

(J) Example cells from (I).

(K) Integration dimension in VMHvl mouse 2.

(L) Same as (I) for VMHvl mouse 2.

(M and N) Same as (K) and (L) for VMHvl mouse

3.

L.W. Swanson
Anatomy of the soul as reflected in the cerebral
hemispheres: neural circuits underlying
voluntary control of basic motivated behaviors
J. Comp. Neurol., 493 (2005), pp. 122-131,
10.1002/cne.20733
View PDF View Record in Scopus
Google Scholar

View in article